Tilde Research introduces Aurora, a leverage-aware optimizer that addresses a hidden neuron death problem in Muon.
Proposes AutoTTS, an environment-driven framework using agentic methods for LLMs to discover and apply test-time scaling strategies, shifting from manual heuristics to automated discovery.
Inside the core ideas, potential and challenges of SSMs
Details Perplexity's inference setup for serving post-trained Qwen3 235B models on NVIDIA Blackwell GPUs, optimizing for cost and performance.
Parameter Golf brought together 1,000+ participants and 2,000+ submissions to explore AI-assisted machine learning research, coding agents, quantization, and novel model design under strict constraints.
Meta, Stanford, and University of Washington researchers propose methods to accelerate Byte Latent Transformer (BLT) generation, reducing inference memory bandwidth by over 50% without tokenization using diffusion and…
Microsoft Research introduced SocialReasoning-Bench, a benchmark evaluating AI agents' social reasoning in calendar coordination and marketplace negotiation, testing outcome optimality and due diligence.
Nimbus builds production AI systems — internal tools, customer agents, retrieval pipelines — combining humans and AI end-to-end. From scoped pilot to production in 4–8 weeks.
Palisade Research shows that AI agents can hack remote computers, copy themselves onto them, and form replication chains. In one year, the success rate jumped from 6 to 81 percent. The researchers expect remaining…
METR conducted risk assessment on an early version of Anthropic's Claude Mythos Preview in March 2026, estimating significant capabilities.
UCLA awarded $5M DARPA grant for ALPHA project to develop AI for automating mathematical proof synthesis and verification in domains like PDEs and number theory.
ACL Anthology PDF version of the work showing autonomous tool creation that can go beyond simple Python functions and produce tools for real-world scientific tasks.